isprs annals IV 4 W1 51 2016

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-4/W1, 2016
1st International Conference on Smart Data and Smart Cities, 30th UDMS, 7–9 September 2016, Split, Croatia

GENERATION OF MULTI-LOD 3D CITY MODELS IN CITYGML WITH THE
PROCEDURAL MODELLING ENGINE RANDOM3DCITY
F. Biljecki ∗, H. Ledoux, J. Stoter
3D Geoinformation, Delft University of Technology, The Netherlands — (f.biljecki, h.ledoux, j.e.stoter)@tudelft.nl

KEY WORDS: Procedural modelling, CityGML, Level of detail (LOD), Multi-Scale

ABSTRACT:
The production and dissemination of semantic 3D city models is rapidly increasing benefiting a growing number of use cases. However,
their availability in multiple LODs and in the CityGML format is still problematic in practice. This hinders applications and experiments
where multi-LOD datasets are required as input, for instance, to determine the performance of different LODs in a spatial analysis. An
alternative approach to obtain 3D city models is to generate them with procedural modelling, which is—as we discuss in this paper—
well suited as a method to source multi-LOD datasets useful for a number of applications. However, procedural modelling has not yet
been employed for this purpose. Therefore, we have developed R ANDOM 3D CITY, an experimental procedural modelling engine for
generating synthetic datasets of buildings and other urban features. The engine is designed to produce models in CityGML and does
so in multiple LODs. Besides the generation of multiple geometric LODs, we implement the realisation of multiple levels of spatiosemantic coherence, geometric reference variants, and indoor representations. As a result of their permutations, each building can be
generated in 392 different CityGML representations, an unprecedented number of modelling variants of the same feature. The datasets
produced by R ANDOM 3D CITY are suited for several applications, as we show in this paper with documented uses. The developed

engine is available under an open-source licence at Github at http://github.com/tudelft3d/Random3Dcity.

1. INTRODUCTION
3D city models are three-dimensional representations of the urban environment. They can be generated with a multitude of approaches: photogrammetry (Suveg and Vosselman, 2004), laser
scanning (Vosselman and Dijkman, 2001, Demir and Baltsavias,
2012), handheld devices (Rosser et al., 2015), conversion from
architectural models (Donkers et al., 2015), radar (Stilla et al.,
2003), and procedural modelling (M¨uller et al., 2006). Each acquisition approach is tied to the level of detail (LOD), a measure
that indicates the spatio-semantic adherence of a model to its realworld equivalent, and the LOD has implications on its usability
(Biljecki et al., 2014b). While in practice the LOD mostly refers
to the richness of the geometry, the concept encompasses also the
granularity of semantics and the amount of attributes (Stadler and
Kolbe, 2007) (see Fig. 1).
Techniques for producing 3D data are capable of deriving data
in multiple LODs (e.g. an airborne laser scanning survey can result in both block models, and models with detailed rooftops).
However, multi-LOD data of the same real world object are still
seldom available, and this will probably not improve in the near
future. This deficiency hinders applications that require them as
input, such as visualisation. There are a few possible reasons
why multi-LOD datasets are rare, among others: (1) GIS software

packages are generally not programmed to utilise multi-scale representations of 3D models; (2) 3D city models are usually acquired for one purpose in mind; hence they are acquired in the
optimal (single) LOD; and (3) there are limitations in the acquisition and storage process, e.g. a 3D modelling software is not capable of simultaneously producing two or more representations,
and to store them consistently and efficiently.

Figure 1: A building modelled in multiple LODs. The LOD
labels are according to the categorisation of CityGML 2.0, and
they denote both the granularity of the geometry and semantics.
Besides small and illustrative models such as this one, building
datasets containing multiple LODs are virtually non-existent.
suited for various applications and experiments. In Section 3, we
introduce R ANDOM 3D CITY, an experimental procedural modelling engine which we have developed to generate buildings in
multiple LODs in the CityGML format. It is the first engine of
this kind, and we have released the code open-source for free
public use. The engine is composed of two modules and it has
been built entirely from scratch with a custom shape grammar.
The datasets generated by this engine have already been used in
several research projects (Section 4).
2. BACKGROUND AND RELATED WORK
2.1 Motivation for multi-LOD models

The aim of this paper is to present our method that we developed
and implemented to generate 3D city models in multiple LODs.
In Section 2 we present related work and we justify procedural
modelling as a potentially suitable acquisition technique to derive such data, which despite their synthetic nature may still be
∗ Corresponding

author

In practice, the vast majority of 3D city models is stored as one
representation, which is sufficient for many single use case scenarios. However, there are situations and use cases in which having multi-LOD data may bring benefit. For instance, (1) visualisation applications and spatial analyses in which, similarly to

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
doi:10.5194/isprs-annals-IV-4-W1-51-2016

51

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-4/W1, 2016
1st International Conference on Smart Data and Smart Cities, 30th UDMS, 7–9 September 2016, Split, Croatia

Design

LOD > i

Planning

Reduction

• Architectural design
• Urban planning
• Procedural modelling

• Generalisation
• Conversion from BIM

Reality

Acquisition
•
•
•

•
•

Photogrammetry
Laser scanning
Surveying
Radar
Architectural drawings

LOD i

Spatial analysis

2

650 kWh/m

• e.g. solar potential

Augmentation

• Extrusion
• Upgrade (adding details)
• Procedural modelling

LOD < i
Figure 2: Life cycle of a 3D city model: different production workflows of 3D city models from the perspective of the level of detail.
computer graphics, multiple LODs are switched for efficient visualisation and to increase cognition, and to decrease computational
complexity (C
¸ o¨ ltekin and Reichenbacher, 2011); (2) as a source
of data for testing software implementations that are focused on
structuring multi-scale data, e.g. compression methods; (3) to enable data producers to differentiate products by stripping down
a high-quality model to attract different segments of the market
(cf. quality discrimination and product versioning); and (4) for
assessing the suitability of a specific LOD prior to tendering and
data acquisition (“LOD benchmarking”) and for serving different applications (i.e. noise models only need block models while
solar potential analysis needs finer detail). For instance, when
planning the procurement of 3D city models for a specific application, it would be beneficial to have a sample dataset in multiple
LODs, run a spatial analysis for each, and compare the accuracy
of the results to their cost of acquisition (for an example see the
paper (Biljecki et al., 2017)). Such performance analysis can then

serve as a decision factor for determining the optimal LOD to be
acquired, prior to acquisition and to avoid procuring data of an
unsuitable LOD (either too fine or too coarse). In this paper we
address the problem of the absence of such experimental datasets.
2.2 CityGML
In our work we focus on the CityGML format, as it is the most
prominent way to store semantic 3D city models. In this way we
also contribute to the growth of publicly available CityGML data.
The OGC CityGML standard represents an information model
and format for storing 3D city models (Gr¨oger and Pl¨umer, 2012,
Open Geospatial Consortium, 2012a, Kolbe, 2009). Among other
strengths, CityGML datasets are semantically structured, enabling
various 3D spatial analyses which require semantic information.
The standard defines five LODs and supports storing data in multiple LODs. For buildings, LOD0 is a 2.5D representation of their
footprints, LOD1 is a coarse prismatic (block) model. LOD2 represents a model with a roof that is modelled as a standardised roof
shape. LOD3 is an architecturally detailed model with openings
such as windows and doors. LOD4 completes an LOD3 by including indoor features (Kolbe, 2009) (see Fig. 1).
CityGML is a relatively new format and despite many advantages
its software support and adoption are still inferior in comparison

to seasoned computer graphics formats such as COLLADA. This
limitation, coupled with the general lack of multi-LOD data, renders multi-LOD CityGML data virtually non-existing in practice
(Biljecki et al., 2015b).
2.3 Overview of acquisition workflows
We have analysed 3D production workflows and accompanying
software support in order to develop an approach to facilitate the
production of multi-LOD data. From the LOD perspective, we
recognise the following groups of approaches to acquire 3D models (Fig. 2):
Direct acquisition The usual approach of deriving 3D city models: a subset of the real-world is abstracted and modelled according to a predefined LOD, e.g. a photogrammetric survey is carried
out to produce LOD2 building models.
Reduction A dataset is obtained by generalisation from an existing 3D city model of a finer LOD. This approach is usually
employed when the 3D model is too complex for the intended
application (Guercke et al., 2011). Generalisation essentially results in multiple representations (the original one and its simplified counterpart).
Augmentation Existing spatial data is used as a base to generate data of a finer LOD. It can be optionally aided by additional
data sources. For instance, if a footprint of a building is available, its height from a cadastral source can be used to generate a
block model (extrusion). Another example is adding roof structures to LOD1 block models to obtain an LOD2 model (Sugihara
and Shen, 2016).
Planning It should also be noted that a fraction of 3D models
are design models which do not represent a real-world setting,
e.g. simulated 3D models used in movies, architectural models

of planned buildings, and models designed by urban planners to
disseminate urban planning concepts (Ben-Joseph et al., 2001).
Such models have not been used much in GIS, but our work
shows that their usability has been underestimated as they can
be used in various GIS experiments where having real-world data
is not essential.

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
doi:10.5194/isprs-annals-IV-4-W1-51-2016

52

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-4/W1, 2016
1st International Conference on Smart Data and Smart Cities, 30th UDMS, 7–9 September 2016, Split, Croatia

When generating data in multiple LODs, the first approach can
be laborious and expensive, and it is hindered by software limitations. Furthermore, the data is burdened with acquisition errors and it may include other inconsistencies such as invalid geometries, an unwanted but common outcome of 3D acquisition
(Ledoux, 2013, Brasebin et al., 2012).
Second, obtaining multi-LOD models with generalisation is viable in theory, however, there is a lack of implemented solutions,
especially those that support CityGML. Furthermore, in this approach a dataset of fine LOD is required, which are usually available for only a limited set of buildings.

2.4 Procedural modelling
In this paper we focus on procedural modelling, as a common
augmentation approach and source of design models, but not previously considered as a source of multi-LOD data. Procedural
modelling involves creating 3D city models from scratch or based
on an existing 2D dataset by using a set of rules (Goetz, 2013,
Martinovi´c, 2015). It is an important topic in computer graphics
and GIS, and there have been several initiatives to develop procedural engines for modelling urban features. For instance, buildings (Wonka et al., 2003, M¨uller et al., 2006, Kelly and Wonka,
2011, Besuievsky and Patow, 2013a), landmarks (Rodrigues et
al., 2010), roads (Beneˇs et al., 2014), plants (Lintermann and
Deussen, 1999), monuments (Koehl and Roussel, 2015), and land
parcels (Vanegas et al., 2012). For a comprehensive list see the
overview in (Smelik et al., 2014).
Procedurally generated models have proven to be an efficient technique for generating 3D models, mostly by enhancing existing
data—e.g. adding fenestration to an LOD2 model (Kim and Wilson, 2014, Tsiliakou et al., 2014, M¨uller Arisona et al., 2013).
They find their use in GIS in an increasing range of applications,
for instance, urban planning and simulation (Besuievsky and Patow, 2014, Rautenbach et al., 2015). An example workflow is
to take block models of buildings, and to synthetically augment
their detail and appearance (e.g. adding a roof shape and textured
fac¸ade) according to a predefined architecture typical for that spatial extent, resulting in a 3D city model with a much finer level of
detail without a significant additional cost. Another example is to

take building footprints as input, e.g. detected from a point cloud
(Hermosilla et al., 2011), and to generate buildings on top of it.
An advantage of this technique is that it is a quick and simple
method to generate 3D city models, usually in large quantities.
Furthermore, the models derived with procedural modelling are
fine in detail (Rodrigues et al., 2010), and because of their nature
they rarely contain topological inconsistencies (e.g. models obtained with automatic reconstruction from LiDAR point clouds
are more susceptible to topological errors).
As a disadvantage, due to their generative nature, procedurally
modelled datasets are not accurate from the GIS point of view: in
fact, they may considerably deviate from the reality they purport
to represent (Musialski et al., 2013). This is due to the primary
goals of procedural modelling: to quickly generate 3D data, and
to increase the LOD of existing models to improve their visual
impression, achieved by artificially adding features. Nevertheless, this inconsistency does not interfere with many applications,
such as flight simulation, and gaming, where the focus is on visualisation, rather than on spatial analyses (Besuievsky and Patow,
2013b). As a result, many synthetic datasets have been generated completely from scratch, representing fictitious settings, for
instance, for movies.
At the moment, the most prominent procedural modelling engine
is ESRI’s CityEngine, widely used by urban planners and other

practitioners. However, so far procedural modelling efforts do not
appear to have been focused much on multi-scale representations
and CityGML. Considering that a procedural engine can be programmed to produce data with a specific granularity, we take advantage of this idea in our work by defining different procedures
to generate buildings in a series of different representations.
3. PROCEDURAL MODELLING ENGINE
RANDOM3DCITY
In this section we present a CityGML compliant procedural modelling engine that we have developed specifically to produce models in multiple LODs. We have formulated and developed a custom methodology, shape grammar, and rules that can be modified
to suit the requirements of a user.
Besides the motivation of tackling the absence of multi-LOD datasets and shortage of diverse sample CityGML data, we have created R ANDOM 3D CITY for other reasons, for instance, to address
the lack of CityGML procedural modelling software to easily create 3D city models in the respective format. The engine has two
functions: generating an unlimited number of synthetic models
that mimic a real-world setting, completely from scratch, and to
augment existing datasets.
The software prototype is composed of two independent and extensible modules (see Figure 3 for the workflow). The first module consists of a customisable set of rules that derives the configuration of the architecture in a parametric description encoded in
an XML format (Section 3.1). This means that the architecture
and rules can be adapted to a specific setting. For instance, different rules can be encoded, such as to imitate a residential area
with buildings that are between 3 and 6 storeys high and have
predominantly flat roofs.
The second module reads the generated parametric data, and realises the parametric description of buildings as 3D city models
in CityGML in multiple representations (Section 3.2).
The module is designed to generate 3D models according to a
refined specification of the LODs of CityGML (Biljecki et al.,
2016a), and generates data in 16 geometric LODs, rather than
only the 5 standard ones. For instance, it creates a variant of
LOD2 with dormers and other roof structures, and another one
without. Besides generating each building in multiple representations distinguished by geometric complexity, the engine generates models in multiple geometric references (e.g. LOD2 with
walls at their actual location and in the other variant with walls
as projections from roof edges; see (Biljecki et al., 2016b) for an
overview); in multiple levels of semantic structuring (e.g. LOD3
with and without the thematically enriched surfaces); with the
difference in geometric type (boundary representations and solids);
and their corresponding indoor geometry further in multiple LODs.
Permuting all these combinations results in 392 representations
of the same building. To the extent of our knowledge, the models obtained with R ANDOM 3D CITY present the most complete
CityGML (and probably in general 3D building) datasets available to date, contributing to a multitude of application domains
(Section 4).
The engine also supports thematic features other than buildings,
and the generation of a basic interior of buildings (Section 3.3).
Each module of the engine is independent, facilitating the extensibility and integration with other data or other engines. For
instance, since the modelled data is first stored in a parametric
form, it can be generated in formats other than CityGML. On the

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
doi:10.5194/isprs-annals-IV-4-W1-51-2016

53

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-4/W1, 2016
1st International Conference on Smart Data and Smart Cities, 30th UDMS, 7–9 September 2016, Split, Croatia

2007, Hammoudi and Dornaika, 2011, Haala and Brenner, 1999,
Henn et al., 2013). They are shown in Figure 5. Furthermore, the
figure demonstrates that a building part such as a garage can also
be generated.

Random3Dcity
Procedural modeler
Randomiser

Rules and
procedures

3D data (CityGML) realisation
Parametric data
(XML)

3D city model
generator

CityGML
datasets

In addition to the geometry and semantic representation, the engine is capable of generating a number of attributes, such as building age, number of floors, use of building, which can be useful for
some use cases. All together, these parameters are stored in an
XML schema (Fig. 6). This example shows the underlying parametric description of the second building from the left in Fig. 5.

Speci!cation (extensible)
Levels of detail
Geometric references
Rules for solids

Figure 3: Modular workflow of the engine R ANDOM 3D CITY.
3.2 CityGML realisation of the parametric building
other hand, an open parametric building format brings two benefits: first, it facilitates the integration with existing data. This is
shown in Section 3.4 where we generate a 3D city model based
on a real-world 2D cadastral dataset, similarly as in present-day
commercial software. And second, the 3D data generator may
be independently used to generate buildings derived from other
sources with the same rules if stored in this form.
Figure 4 illustrates an example of the output of the engine: a
setting with 100 synthetic buildings in four LODs which are randomly placed and rotated. All datasets (incl. the 388 others not
shown here) have been generated in less than one minute.
The software has been implemented in Python, and it is available
as open-source at Github. A set of generated test datasets is also
available for public use on the website of the project.
3.1 Parametric description and rules
In the engine, each building and other real-world feature Fi consists of a set of n parameters pi that define its architecture:
Fi = {p1i , p2i , . . . pn
i }

(1)

These parameters are in line with the ones used in photogrammetry (e.g. cf. (Zhang et al., 2014, Haala and Kada, 2010, SinningMeister et al., 1996)), such as the width of a building, length of
ridges and eaves, and roof height. The first part of the engine,
described in this section, creates features described by these parameters, using an encoded grammar and a set of rules. The parameters are then stored in an XML schema that we have defined.
Our methodology of procedurally modelling the urban features
consists of first defining a top-down hierarchy of parameters pi .
For instance, before determining the number of windows on a
wall, first the width and length of the footprint of the building
are derived. Thanks to the hierarchical approach, the parameters are context-aware, i.e. the engine contains several procedural rules and constraints. For instance, flat roofs cannot contain
dormers, doors have to be located on the ground floor, buildings
have to have at least one floor, and windows cannot be taller than
the height of the floor. Second, a range for each parametre pi
is defined, e.g. the width of the window is between 0.5 and 1.5
m, from where the engine randomly samples a value according
to a defined probability distribution function (also customisable).
Third, the rules take care to generate a realistic setting, for instance, that multiple dormers are properly aligned on the roof.
All these rules are stored in the code in a series of IF-THEN statements, and they are customisable to conform to a specific setting
a user aims to (re)construct.
R ANDOM 3D CITY supports five types of roofs, which are frequently described as the most common types of roofs (Kada,

The second part of our engine reads the generated parametric data
pi of each feature Fi and constructs CityGML 2.0 datasets in
multiple LODs. The process of the construction of the geometric
representation from the parametric representation is described in
this section by highlighting a few important aspects.
3.2.1 Construction of the geometry and the CityGML file
Each of the representations is generated separately. For each, the
engine reads only the set Pi of parameters pi that it requires for
constructing it. For instance, for the LOD2, a set PiLOD2 is defined and the engine fetches the size of the body of the building,
roof type, and height of the roof, ignoring other parameters such
as windows and dormers. These instructions are stored in the engine, and can be customised to create additional custom LODs,
if required. For instance, it is possible to generate an LOD3-like
model that contains only roof openings, by simply disabling the
constructor for other features such as dormers.
In the construction of the geometry, the engine first creates a local Cartesian coordinate system Xi for the building Fi , and then a
coordinate system Xpi for each of its elements defined by pi (e.g.
wall). The vertices of the features are generated in this system
(e.g. origin of the window on a wall), and the resulting surfaces
are generated. For each different element class (e.g. chimney), an
algorithm is designed for their geometric realisation from their
parametric description. These “sub-systems” Xpi are then converted to the coordinate system Xi of the building, which is later
converted in the global system defined by the user, depending on
the location and orientation of the building in space.
The process of the generation of the vertices of the building elements is not equal for all buildings, it mostly depends on the type
of the roof. For instance, the vertices of the walls are not equal for
buildings with a flat and gabled roof (visible in Figure 5), hence,
separate algorithms are designed for each roof type.
In the process of the generation, geometries are given a universally unique identifier (UUID) according to (ISO, 2008), the recommended approach from CityGML and GML (Open Geospatial
Consortium, 2012b), and are then structured according to the semantic level of the representation. Furthermore, the attributes
have been translated and stored according to the CityGML 2.0
standard. An excerpt of the CityGML of the building exemplified
through its parametric description in the previous section is given
in Figure 8, which also shows the realised attributes.
3.2.2 Generation of corresponding solids Each model, with
the exception of LOD0 models, is also stored as a gml:Solid.
Solids are used to facilitate uses in application domains which
require the usable volume of a building, such as for the estimation
of property taxes (Boeters et al., 2015), and the estimation of the
energy demand of households (Strzalka et al., 2011).

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
doi:10.5194/isprs-annals-IV-4-W1-51-2016

54

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-4/W1, 2016
1st International Conference on Smart Data and Smart Cities, 30th UDMS, 7–9 September 2016, Split, Croatia

Figure 4: Composite rendering of a few exemplary datasets generated by R ANDOM 3D CITY: randomly generated buildings realised in
CityGML in four LODs. The two representations on the left have their walls modelled as projections from roof edges as opposed to
the third panel where the walls are modelled at their actual location and where roof overhangs are explicitly modelled, showing varying
geometric references that are supported by the engine, besides multiple geometric and semantic LODs.
Figure 6: Example of the building parameters stored in an XML.

Figure 5: Different types of roofs (flat, gabled, hipped, pyramidal, and shed) fitted on the same building, with the automatically
adjusted walls. This image shows the LOD2 (orthogonal view),
and it also features an example of a building part.
In the construction of solids, features that do not contribute towards the usable volume of buildings are disregarded. This applies for instance to roof overhangs, and chimneys. In the gml:
MultiSurface representation this is also regarded by the generation of a ClosureSurface to seal the open sides and to provide
a representation as geometrically closed volume object, following
the recommendation of (Gr¨oger and Pl¨umer, 2012).
Figure 9 shows an example of the relation between the semantic
boundary representation models and its solid counterpart. The
solids generated by the engine have been geometrically validated
according to the standard ISO 19107 (ISO, 2003) with the implementation of (Ledoux, 2013).
3.3 Generation of indoor and non-building features
We have implemented multiple versions of the interior according
to the refinement developed by (Boeters et al., 2015). Further,
thematic features other than buildings have been generated, such
as vegetation and roads. Figure 7 shows an example with the
interior of buildings (LOD2+ model as per (Boeters et al., 2015)).
3.4 Augmenting existing data (2D footprints)
Besides generating all parameters of features from scratch, an
experimental feature of the engine is to utilise existing GIS data

< building ID = " c209f43f - f137 -4 dd0 -8816 - cbc4dc4a407d " >
< footprint > Rectangular
< origin > 173469.34 526427.95 0.0
< rotation > 34.3
< xSize > 7.06
< ySize > 8.91
< zSize > 6.8
< floors >2
< floorHeight > 3.4
< embrasure > 0.08
< wallThickness > 0.2
...
< buildingPart >
< partType > Garage
...

< roof >
< roofType > Gabled
2.48
< overhangs >
< xlength > 0.5
< ylength > 0.5
...

such as footprints, and procedurally model the remaining features
producing a 3D model. For instance, Figure 10 shows a setting in
which 2D footprints from an existing dataset have been used, and
the 3D buildings have been procedurally modelled by specifying
the architecture where for instance only hipped roofs are present
and where buildings must have large windows (as it is the case
for that setting).
4. DOCUMENTED APPLICATIONS OF OUR
SOFTWARE
Availability of freely available datasets with a large number of
dissimilar buildings represented in multiple LODs opens a door

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
doi:10.5194/isprs-annals-IV-4-W1-51-2016

55

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-4/W1, 2016
1st International Conference on Smart Data and Smart Cities, 30th UDMS, 7–9 September 2016, Split, Croatia

Figure 7: A composite street view of a dataset generated by the engine, showing the vegetation (park), street network, and the basic
indoor representation of buildings (a solid representing each storey).
< cityObjectMember >
< bldg : Building gml : id = " c209f43f - f137 4 dd0 -8816 - cbc4dc4a407d " >
< bldg : roofType > Gabled

< bldg : y e a r O f C o n s t r u c t i o n > 2008

< bldg : s t o r e y s A b o v e G r o u n d >2

< bldg : boundedBy >
< bldg : GroundSurface >
...

Figure 8: Excerpt of the generated CityGML 2.0 dataset.
for a multitude of research purposes. Instead of an experiment
section, we showcase the documented uses of preliminary versions of R ANDOM 3D CITY from fellow researchers in the 3D GIS
community. The generated datasets have already been tested and
used in several application domains:

Figure 9: Two representations of LOD3 models: a gml:Solid
and a thematically structured gml:MultiSurface. Note the
ClosureSurface of the chimney (light brown) in order to separate the usable volume of a building.
rameters by sampling values from a normal probability distribution function with standard deviation simulating acquisition errors. This approach results in ground truth and erroneous versions of parameters, which the second part of the
engine used as an input to create reference and erroneous 3D
city models, suited for uncertainty propagation experiments.
This is the first instance in which procedurally generated 3D
data has been used for uncertainty propagation research.

1. Testing and improving CityGML validation software (Ledoux,
2013, Zhao et al., 2013, Coors and Wagner, 2015), primarily
in the scope of the OGC CityGML Quality Interoperability Experiment (QIE) (OGC, 2016). Considering that the
engine creates topologically consistent datasets, they have
been used as exemplary models of valid datasets. Furthermore, for this project the code of the engine was modified
to intentionally produce data with topological errors, such as
overlapping buildings and broken solids, which were used as
input to test validation and repair software packages benchmarked in the QIE.

3. Optimising the coverage of geosensor networks (Doodman
et al., 2014, Afghantoloee et al., 2014). Researchers have
used our synthetic datasets to test the implementation of
their use case, which focuses on line of sight analysis.

2. Analyse the propagation of uncertainty from the input data
to the result of a GIS operation with a Monte Carlo simulation (Biljecki et al., 2014a, Biljecki et al., 2015a). A stochastic engine on top of the first module of R ANDOM 3D CITY
was implemented to intentionally degrade the building pa-

4. For testing software which identifies and links topological
relationships between similar features across multiple LODs
in CityGML (Biljecki et al., 2015b). The project focused on
improving the consistency of multi-LOD datasets and their
compression. R ANDOM 3D CITY was found valuable in this

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
doi:10.5194/isprs-annals-IV-4-W1-51-2016

56

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-4/W1, 2016
1st International Conference on Smart Data and Smart Cities, 30th UDMS, 7–9 September 2016, Split, Croatia

Obviously, this experimental software cannot compete with advanced commercial solutions, such as ESRI’s CityEngine, which
are capable of creating complex architecture. Nevertheless, it
bridges the gap with respect to CityGML data and multiple representations to quickly obtain models suited for experiments and
testing, and it is open-source.

Figure 10: Example of a CityGML model generated with R AN DOM 3D CITY in conjunction with the existing real-world dataset
of 2D footprints of buildings. This setting shows the Beestenmarkt in Delft, the Netherlands. The grammar has been adjusted
to match the configuration of the buildings in that setting in order
to resemble the reality as close as possible.
project as it presented the only source of such data.
5. As a data source in experiments with voxelisation of CityGML
models to facilitate volume computation (Steuer et al., 2015).
6. To test the specification of LODs and their differences when
used in a spatial analysis (Biljecki et al., 2016a, Biljecki et
al., 2017). In these projects, several LODs have been benchmarked in different spatial analyses, such as estimating the
shadow cast by a building, to assess if a finer level of detail
brings an improvement in a particular spatial analysis.
5. CONCLUSIONS AND FUTURE PROSPECTS
In this paper we have tackled the absence of multi-LOD 3D city
models by exploring the possibility of generating them with procedural modelling, as opposed to orthodox techniques such as
laser scanning and generalisation. We have developed a procedural modelling engine that is designed with the primary objective
to provide models in multiple LODs stored in the OGC CityGML
format. In this way we also address the lack of procedural modelling engines that support CityGML, and the shortage of publicly
available CityGML data.
The experimental engine R ANDOM 3D CITY, which we have introduced and built from scratch, is novel: it natively supports
CityGML, and it is designed towards producing multi-LOD data.
It yields an unprecedented number of variants of models, and
does so according to an underlying set of customisable rules. The
engine is open-source, and a set of example datasets is available for free public use. The reason why we have made this
project open is that other researchers can benefit from multi-LOD
datasets in their application domains, and that they could adjust
its grammar to suit specific requirements. As a result, the generated data have already been proven useful by being featured
in several research projects in different countries and application
domains. Such interest suggests the need for open procedural
modelling engines. We invite other researchers to take advantage
of the availability of these datasets to test their implementations
and as a source for experiments.
While the engine generates fictitious settings, it is still suited
for applications where having real-world data is not important,
and where different scenarios can be evaluated (e.g. to determine
whether it is more beneficial to acquire an LOD2 instead of an
LOD1 for a specific spatial analysis, by testing both representations before the actual acquisition).

For future work we plan to work in two directions. First, we
plan to advance the shape grammar for generating more complex
buildings, such as structures with less usual roof types and landmarks. This is a natural flow of the work, which is hampered by
designing advanced rules that more complex features imply, such
as complicated roof shapes and balconies. Second, we plan to
increase the LOD of the interior. We have introduced the modelling of a basic indoor (storeys), and we intend to include the
generation of rooms and openings (windows/doors), following
recent efforts in procedural modelling of interior according to an
indoor grammar (Becker et al., 2013, Peter et al., 2013, Gr¨oger
and Pl¨umer, 2010, Ilˇc´ık and Wimmer, 2013).
ACKNOWLEDGEMENTS
We gratefully acknowledge the comments of the anonymous reviewers, and of Micka¨el Brasebin (IGN France) during the development phase. This research is supported by the Dutch Technology Foundation STW, which is part of the Netherlands Organisation for Scientific Research (NWO), and which is partly funded
by the Ministry of Economic Affairs (project code: 11300).
REFERENCES
Afghantoloee, A., Doodman, S., Karimipour, F. and Mostafavi, M. A.,
2014. Coverage estimation of geosensor in 3D vector environments. Int.
Arch. Photogramm. Remote Sens. Spatial Inf. Sci. XL-2/W3, pp. 1–6.
Becker, S., Peter, M., Fritsch, D., Philipp, D., Baier, P. and Dibak, C.,
2013. Combined Grammar for the Modeling of Building Interiors. ISPRS
Ann. Photogramm. Remote Sens. Spatial Inf. Sci. II-4/W1, pp. 1–6.
Ben-Joseph, E., Ishii, H., Underkoffler, J., Piper, B. and Yeung, L., 2001.
Urban simulation and the luminous planning table: bridging the gap between the digital and the tangible. Journal of Planning Education and
Research 21(2), pp. 196–203.
Beneˇs, J., Wilkie, A. and Kˇriv´anek, J., 2014. Procedural Modelling of
Urban Road Networks. Computer Graphics Forum 33(6), pp. 132–142.
Besuievsky, G. and Patow, G., 2013a. Customizable LoD for Procedural
Architecture. Computer Graphics Forum 32(8), pp. 26–34.
Besuievsky, G. and Patow, G., 2013b. Procedural modeling historical
buildings for serious games. Virtual Archeology Review 4(9), pp. 160–
166.
Besuievsky, G. and Patow, G., 2014. Recent Advances on LoD for Procedural Urban Models. In: Proceedings of the Workshop on Processing
Large Geospatial Data, Cardiff, United Kingdom.
Biljecki, F., Heuvelink, G. B. M., Ledoux, H. and Stoter, J., 2015a. Propagation of positional error in 3D GIS: estimation of the solar irradiation
of building roofs. International Journal of Geographical Information Science 29(12), pp. 2269–2294.
Biljecki, F., Ledoux, H. and Stoter, J., 2014a. Error propagation in the
computation of volumes in 3D city models with the Monte Carlo method.
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci. II-2, pp. 31–39.
Biljecki, F., Ledoux, H. and Stoter, J., 2015b. Improving the consistency of multi-LOD CityGML datasets by removing redundancy. In: 3D
Geoinformation Science, Springer International Publishing, pp. 1–17.
Biljecki, F., Ledoux, H. and Stoter, J., 2016a. An improved LOD specification for 3D building models. Computers, Environment and Urban
Systems 59, pp. 25–37.
Biljecki, F., Ledoux, H. and Stoter, J., 2017. Does a finer level of detail
of a 3D city model bring an improvement for estimating shadows? In:
Advances in 3D Geoinformation, Springer International Publishing.

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
doi:10.5194/isprs-annals-IV-4-W1-51-2016

57

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-4/W1, 2016
1st International Conference on Smart Data and Smart Cities, 30th UDMS, 7–9 September 2016, Split, Croatia

Biljecki, F., Ledoux, H., Stoter, J. and Vosselman, G., 2016b. The variants
of an LOD of a 3D building model and their influence on spatial analyses.
ISPRS Journal of Photogrammetry and Remote Sensing 116, pp. 42–54.

Kim, K. and Wilson, J. P., 2014. Planning and visualising 3D routes for
indoor and outdoor spaces using CityEngine. Journal of Spatial Science
60(1), pp. 179–193.

Biljecki, F., Ledoux, H., Stoter, J. and Zhao, J., 2014b. Formalisation
of the level of detail in 3D city modelling. Computers, Environment and
Urban Systems 48, pp. 1–15.

Koehl, M. and Roussel, F., 2015. Procedural modelling for reconstruction
of historic monuments. ISPRS Ann. Photogramm. Remote Sens. Spatial
Inf. Sci. II-5-W3, pp. 137–144.

Boeters, R., Arroyo Ohori, K., Biljecki, F. and Zlatanova, S., 2015. Automatically enhancing CityGML LOD2 models with a corresponding indoor geometry. International Journal of Geographical Information Science 29(12), pp. 2248–2268.

Kolbe, T. H., 2009. Representing and exchanging 3D city models with
CityGML. In: 3D Geo-Information Sciences, Springer Berlin Heidelberg,
pp. 15–31.

Brasebin, M., Perret, J., Musti`ere, S. and Weber, C., 2012. Measuring the
impact of 3D data geometric modeling on spatial analysis: Illustration
with Skyview factor. In: Usage, Usability, and Utility of 3D City Models – European COST Action TU0801, EDP Sciences, Nantes, France,
pp. (02001)1–16.
C
¸ o¨ ltekin, A. and Reichenbacher, T., 2011. High Quality Geographic Services and Bandwidth Limitations. Future Internet 3(4), pp. 379–396.
Coors, V. and Wagner, D., 2015. CityGML Quality Interoperability Experiment des OGC. DGPF Tagungsband. Publikationen der Deutschen
Gesellschaft f¨ur Photogrammetrie, Fernerkundung und Geoinformation
e.V. 24, pp. 288–295.
Demir, N. and Baltsavias, E. P., 2012. Automated modeling of 3D building roofs using image and LiDAR data. ISPRS Ann. Photogramm. Remote
Sens. Spatial Inf. Sci. I-4, pp. 35–40.
Donkers, S., Ledoux, H., Zhao, J. and Stoter, J., 2015. Automatic conversion of IFC datasets to geometrically and semantically correct CityGML
LOD3 buildings. Transactions in GIS.
Doodman, S., Afghantoloee, A., Mostafavi, M. A. and Karimipour, F.,
2014. 3D extension of the vor algorithm to determine and optimize the
coverage of geosensor networks. Int. Arch. Photogramm. Remote Sens.
Spatial Inf. Sci. XL-2/W3, pp. 103–108.
Goetz, M., 2013. Towards generating highly detailed 3D CityGML models from OpenStreetMap. International Journal of Geographical Information Science 27(5), pp. 845–865.

Ledoux, H., 2013. On the Validation of Solids Represented with the International Standards for Geographic Information. Computer-Aided Civil
and Infrastructure Engineering 28(9), pp. 693–706.
Lintermann, B. and Deussen, O., 1999. Interactive modeling of plants.
IEEE Computer Graphics and Applications 19(1), pp. 56–65.
Martinovi´c, A., 2015. Inverse Procedural Modeling of Buildings. PhD
thesis, KU Leuven.
M¨uller Arisona, S., Zhong, C., Huang, X. and Qin, H., 2013. Increasing
detail of 3D models through combined photogrammetric and procedural
modelling. Geo-spatial Information Science 16(1), pp. 45–53.
M¨uller, P., Wonka, P., Haegler, S., Ulmer, A. and van Gool, L., 2006.
Procedural modeling of buildings. ACM Transactions on Graphics 25(3),
pp. 614–623.
Musialski, P., Wonka, P., Aliaga, D. G., Wimmer, M., van Gool, L. and
Purgathofer, W., 2013. A Survey of Urban Reconstruction. Computer
Graphics Forum 32(6), pp. 146–177.
OGC, 2016. OGC CityGML Quality Interoperability Experiment. Technical Report OGC 16-064.
Open Geospatial Consortium, 2012a. OGC City Geography Markup Language (CityGML) Encoding Standard 2.0.0. Technical report.
Open Geospatial Consortium, 2012b. OGC Geography Markup Language (GML) — Extended schemas and encoding rules 3.3.0.
Peter, M., Becker, S. and Fritsch, D., 2013. Grammar supported indoor
mapping. In: Proceedings of the 26th International Cartographic Conference, Dresden, Germany, pp. 1–18.

Gr¨oger, G. and Pl¨umer, L., 2010. Derivation of 3D Indoor Models by
Grammars for Route Planning. Photogrammetrie - Fernerkundung Geoinformation 2010(3), pp. 191–206.

Rautenbach, V., Bevis, Y., Coetzee, S. and Combrinck, C., 2015. Evaluating procedural modelling for 3D models of informal settlements in urban
design activities. South African Journal of Science 111(11/12), pp. 1–10.

Gr¨oger, G. and Pl¨umer, L., 2012. CityGML – Interoperable semantic 3D
city models. ISPRS Journal of Photogrammetry and Remote Sensing 71,
pp. 12–33.

Rodrigues, R., Coelho, A. and Reis, L. P., 2010. Data model for procedural modelling from textual descriptions. In: IEEE Congress on Evolutionary Computation (CEC), IEEE, Barcelona, Spain, pp. 1–8.

Guercke, R., G¨otzelmann, T., Brenner, C. and Sester, M., 2011. Aggregation of LoD 1 building models as an optimization problem. ISPRS
Journal of Photogrammetry and Remote Sensing 66(2), pp. 209–222.

Rosser, J., Morley, J. and Smith, G., 2015. Modelling of Building Interiors with Mobile Phone Sensor Data. ISPRS International Journal of
Geo-Information 4(2), pp. 989–1012.

Haala, N. and Brenner, C., 1999. Virtual city models from laser altimeter
and 2D map data. Photogrammetric Engineering and Remote Sensing
65(7), pp. 787–795.

Sinning-Meister, M., Gruen, A. and Dan, H., 1996. 3D city models for
CAAD-supported analysis and design of urban areas. ISPRS Journal of
Photogrammetry and Remote Sensing 51(4), pp. 196–208.

Haala, N. and Kada, M., 2010. An update on automatic 3D building
reconstruction. ISPRS Journal of Photogrammetry and Remote Sensing
65(6), pp. 570–580.

Smelik, R. M., Tutenel, T., Bidarra, R. and Benes, B., 2014. A Survey
on Procedural Modelling for Virtual Worlds. Computer Graphics Forum
33(6), pp. 31–50.

Hammoudi, K. and Dornaika, F., 2011. A Featureless Approach to
3D Polyhedral Building Modeling from Aerial Images. Sensors 11(1),
pp. 228–259.

Stadler, A. and Kolbe, T. H., 2007. Spatio-semantic coherence in the integration of 3D city models. Int. Arch. Photogramm. Remote Sens. Spatial
Inf. Sci. XXXVI-2/C43, pp. 8.

Henn, A., Gr¨oger, G., Stroh, V. and Pl¨umer, L., 2013. Model driven
reconstruction of roofs from sparse LIDAR point clouds. ISPRS Journal
of Photogrammetry and Remote Sensing 76, pp. 17–29.

Steuer, H., Machl, T., Sindram, M., Liebel, L. and Kolbe, T. H., 2015.
Voluminator—Approximating the Volume of 3D Buildings to Overcome
Topological Errors. In: AGILE 2015, Springer International Publishing,
pp. 343–362.

Hermosilla, T., Ruiz, L. A., Recio, J. A. and Estornell, J., 2011. Evaluation of Automatic Building Detection Approaches Combining High Resolution Images and LiDAR Data. Remote Sensing 3(6), pp. 1188–1210.
Ilˇc´ık, M. and Wimmer, M., 2013. Challenges and ideas in procedural
modeling of interiors. In: Eurographics Workshop on Urban Data Modelling and Visualisation, Girona, pp. 29–30.
ISO, 2003. 19107:2003(E) – Geographic information - Spatial schema.
ISO, 2008. ISO/IEC 9834-8:2008(E) – Information technology — Open
Systems Interconnection — Procedures for the operation of OSI Registration Authorities: Generation and registration of Universally Unique
Identifiers (UUIDs) and their use as ASN.1 Object Identifier components.
Kada, M., 2007. Scale-Dependent Simplification of 3D Building Models Based on Cell Decomposition and Primitive Instancing. In: Spatial
Information Theory, Springer Berlin Heidelberg, pp. 222–237.
Kelly, T. and Wonka, P., 2011. Interactive architectural modeling with
procedural extrusions. ACM Transactions on Graphics 30(2), pp. 1–15.

Stilla, U., Soergel, U. and Thoennessen, U., 2003. Potential and limits of
InSAR data for building reconstruction in built-up areas. ISPRS Journal
of Photogrammetry and Remote Sensing 58(1-2), pp. 113–123.
Strzalka, A., Bogdahn, J., Coors, V. and Eicker, U., 2011. 3D City modeling for urban scale heating energy demand forecasting. HVAC&R Research 17(4), pp. 526–539.
Sugihara, K. and Shen, Z., 2016. Automatic generation of 3D house
models with solar photovoltaic generation for smart city. In: 3rd MEC
International Conference on Big Data and Smart City (ICBDSC), Muscat,
Oman, pp. 1–7.
Suveg, I. and Vosselman, G., 2004. Reconstruction of 3D building models
from aerial images and maps. ISPRS Journal of Photogrammetry and
Remote Sensing 58(3-4), pp. 202–224.
Tsiliakou, E., Labropoulos, T. and Dimopoulou, E., 2014. Procedural
Modeling in 3D GIS Environment. International Journal of 3-D Information Modeling 3(3), pp. 17–34.

This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
doi:10.5194/isprs-annals-IV-4-W1-51-2016

58

ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume IV-4/W1, 2016
1st International Conference on Smart Data and Smart Cities, 30th UDMS, 7–9 September 2016, Split, Croatia

Vanegas, C. A., Kelly, T., Weber, B., Halatsch, J., Aliaga, D. G. and
M¨uller, P., 2012. Procedural Generation of Parcels in Urban Modeling.
Computer Graphics Forum 31(2), pp. 681–690.
Vosselman, G. and Dijkman, S., 2001. 3D building model reconstruction
from point clouds and ground plans. Int. Arch. Photogramm. Remote
Sens. Spatial Inf. Sci. XXXIV-3/W4, pp. 37–44.
Wonka, P., Wimmer, M., Ribarsky, W. and Sillion, F., 2003. Instant architecture. ACM Transactions on Graphics 22(3), pp. 669–677.

Zhang, W., Wang, H., Chen, Y., Yan, K. and Chen, M., 2014. 3D Building
Roof Modeling by Optimizing Primitive’s Parameters Using Constraints
from LiDAR Data and Aerial Imagery. Remote Sensing 6(9), pp. 8107–
8133.
Zhao, J., Ledoux, H. and Stoter, J., 2013. Automatic repair of CityGML
LoD2 buildings using shrink-wrapping. ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci. II-2/W1, pp. 309–317.

This contributio